This study makes a useful contribution to studying rain-on-events (at ephemeral snowpacks) by developing and evaluating an algorithm that can detect basin-scale rain-on-snow events using gridded climate data. The method searches for periods of concurrent (area-averaged) precipitation, surface runoff, and snowmelt exceeding pre-defined values. Application of this to Susquehanna River Basin (SRB) indicates that (using dataset-specific thresholds) the method seems to work appropriately. The method also highlights there can be large differences in RoS event magnitude and frequency caused by differences in the various driving factors. 

In principle, this paper could be published, as the developments seem useful, and are logically shown. However, at the same time, the sort of ad-hoc presentation of a single basin results makes it unclear how useful the methods are, and how it’s usefulness varies between scales and settings. This would make the paper more useful and thereby a better fit for this journal.

• It would be useful if a clearer workflow is presented. This helps other people to better understand and potentially adopt the approach 
• The method is applied to one basin. This is OK, but also a bit a thin basis for introducing a method. Across what range of areas is your method likely applicable?
• The comparison of climate data for a single basin (section 3.1) is useful, but it would be a lot more useful to systematically test how these things vary? (i.e. beyond this specific basin).  
• The timescale over which RoS are relevant will be very dependent on the spatial scale of the watershed. For example, headwater catchments may need timescales of hours whereas for large basins it may be multiple days. Can this be discussed?
• Provide units at all axes go all figures. 

The authors present an inter-model comparison of basin-scale rain-on-snow (RoS) flood events identified using different detection algorithms for a 21-year historical period over the Susquehanna River Basin, located in the eastern US. Compared to observations, the algorithm flags known historical events, but there are model-specific discrepancies in terms of event magnitude and the driving factors of precipitation, runoff, snowpack, and snowmelt contribution. The authors report that the use of dataset-specific thresholds improves agreement between models but does not account for all discrepancies, which are attributed to differences in model structure, meteorological forcing, and coupling frequency.

The paper is well-written and clearly structured. The references cited are complete and up to date. The results are relevant to local and regional flood risk assessments using gridded climate and hydrometeorological data sets that are available at continental to global scales.

I have two general comments and a few detailed edits.

Comment #1:

It is mentioned that the ROF differences among hydrologic models can be extremely variable, with regional differences between products reaching an order of magnitude. Is this due to soil parameterization in models that influences infiltration capacity? Is there a difference in soil parameterization that sets L15 apart from the other models? Given the high uncertainty in soil parameterization, I wonder if there is benefit in including ROF as a threshold.

There is a misunderstanding evident on Line 125. The statement “20% reduction in the snowpack (implied dSWE <=  -20 mm/day)” is incorrect. The 20% snowmelt contribution threshold applied by Freudiger et al. (2014), which was subsequently adopted by Musselman et al. (2018) and Li et al. (2019), is described as: “… for which the sum of rainfall and snowmelt contains a minimum of 20% snowmelt.” This is a key point to make clear. What is helpful about this threshold condition is that it doesn’t rely on characterization of soil infiltration capacity, which is difficult to validate in models and can vary over an order of magnitude regionally and between models.

What I like about your inclusion of the ROF variable is that it connects to the importance of antecedent soil moisture and saturated conditions to driving flood response. More discussion on this topic is probably warranted. Requiring ROF>0 is a condition that the soil must be saturated to result in flood, which is intuitive, and then not surprising that the flagged events in the simulations correspond to streamflow events in the USGS record.

Given that (at least) these three papers adopted the “20%” threshold to identify times when snowmelt is substantially contributing to rainfall + snowmelt fluxes, I encourage the authors to include this 20% value in their FIXED (and X% in their RELATIVE) analyses. Given the uncertainty associated with estimating ROF, can you identify flood events in the record without requiring ROF? What is potentially missed / lost if one doesn’t include ROF? For example, excluding ROF as a condition is analogous to not knowing the buffering capacity of soil water deficit. Would that impact the ability of an automated algorithm to flag potentially impactful RoS events? This would seem to be an important tie to previous studies – because I do think antecedent soil saturation is important - and could result in key guidance to future research efforts.

Comment #2:

It would be helpful to see more written about the transferability of the results to other regions / basins. Particularly, how transferrable would your results be to regions with more elevation complexity as it relates to model treatment of elevation-dependent hydrometeorology, snowpack, and soil antecedent conditions, which would seem relevant to the differences in model grid spacing? If not transferrable, could the authors provide guidance for future research efforts?

Detailed Edits:

Line 181: “(i.e., positive dSWE in both figures denotes snowpack)”, I think “denotes snow accumulation” is more accurate.

Line 225: This should be “WY 1996”. In the US, the water year is designated by the calendar year in which it ends. The year ending September 30, 1996 is called the "1996 WY”, which begins on Oct 1, 1995.

Figures 2 & 3: The histograms in panels b-d are very difficult for me to visually disentangle. Please consider making each variable-panel into 2x2 sub-panels showing the individual (solid / filled) colored histograms for each model.  

Thank you for the chance to review this important work.

Sincerely,

Keith Musselman